An artificial intelligence model to predict upcoming embryology workload for patients undergoing ovarian stimulation

Background: The daily workload for an embryology laboratory varies depending on the number of egg retrievals performed each day and how many eggs were retrieved from each patient. An artificial intelligence (AI) model to predict upcoming embryology workload could assist with staffing and identify which days are likely to necessitate a higher resource allocation for procedures such as intracytoplasmic sperm injection (ICSI) and preimplantation genetic testing (PGT).

Objective: To develop an AI model that, for any given day, evaluates all patients currently undergoing ovarian stimulation and predicts the timing of egg retrievals and total number of mature eggs (MIIs) per day up to 9 days in the future.

Materials and Methods: Historical, de-identified electronic medical record (EMR) data were collected for IVF retrieval cycles started between 2014-2020 from 4 different sized IVF clinics in the United States, for a total of approximately 37,500 IVF cycles. EMR records contained patient baseline characteristics, daily measurements of estradiol (E2), and daily follicle counts and sizes. Patient cycles were split into train and test datasets stratified by site. A first CatBoost regression model was trained to predict the number of MIIs a patient would retrieve using patient age, AMH, cycle day, and follicle and E2 growth trends. A second CatBoost regression model was trained using the same input parameters to predict a patient’s remaining cycle duration. Predictions were fitted with a probability density function and optimized for calibration accuracy. Together, these models provide a prediction, for any given cycle day, of when a patient is likely to be triggered and how many MIIs they will have. The AI model was tested by evaluating all patients undergoing stimulation on a given day and predicting the total number of MIIs per day up to 9 days into the future.

Results: On the test dataset comprising 25% of the patient cycles, the MII prediction model had a mean absolute error (MAE) of 3.21 eggs and an R² of 0.58. The cycle duration prediction model had an MAE of 0.65 days and an R² of 0.88. Using increments of 1 week across the entire test dataset and predicting the total number of MIIs per day up to 9 days in the future, the combined model had an R² of 0.66. Furthermore, the model identified future high ICSI workload days (defined as MIIs exceeding the 75th percentile) with an accuracy of 86% and a weighted F1 score of 0.86.

Conclusions: We developed an AI model for forecasting embryology laboratory workload up to 9 days in the future for patients undergoing ovarian stimulation. This tool could help identify high workload days and help with staffing and resource allocation.